Fix audio placeholder length mismatch for merged audio inputs
@cherry77-cloud
Thanks for the detailed report and repro script — really appreciate the thorough investigation.
I think the issue here isn't a code bug, but rather that merge_audio_from_same_content=True may not be the right fit for this use case.
This parameter was designed for simplex omni streaming, where inputs are fixed 1-second audio chunks. At 16kHz, 1s = 16000 samples divides cleanly by hop_length=160, so F(16000) + F(16000) = F(32000) — the placeholder and embedding counts always align.
In your case, each audio clip is 33440 samples, which doesn't divide evenly by 160. That introduces a rounding mismatch: F(33440) + F(33440) ≠ F(66880), giving 42 placeholders but only 41 embeddings from the merged waveform.
Simply switching to merge_audio_from_same_content=False should resolve it — in that mode both placeholder calculation and audio encoding process each clip independently, so they'll always match. No code changes needed.
Hope that helps!
@tc-mb 。thank you so much!
You're absolutely right — this isn't a bug at all, and I should have read the
documentation more carefully before filing the report.
Really appreciate your patience and the detailed breakdown.